DDFusion: An efficient multi-exposure fusion network with dense pyramidal convolution and de-correlation fusion.

In this work, we propose DDFusion, a novel multi-exposure image fusion network. DDFusion addresses the limitations of existing methods by effectively recovering details near extremely bright regions and learning associations between non-contiguous regions. To achieve this, our network incorporates a dense pyramidal (DensePy) convolution block in the encoder for multi-scale feature extraction, and a de-correlation fusion (DF) block for enabling structurally coherent and edge-preserving multi-scale feature fusion. It facilitates a smoother transition from highlighted areas to adjacent regions in the fused image. Experimental results demonstrate the superiority of DDFusion over state-of-the-art deep methods in terms of both visual quality and quantitative evaluation. Moreover, DDFusion achieves stronger multi-scale feature extraction capability with smaller computational complexity. • The proposed dense pyramidal convolution block combines PyConv and dense connections, which forms a multi-scale feature expression with more continuous in scale and complementary in content. • The designed DF block contains FastDeconv and an attention module, which enhances high-frequency edge details and fuses multi-scale structural features smoothly, enabling structurally coherent and edge-preserving multi-scale feature fusion. • The proposed DDFusion integrates dense pyramidal convolution and DF blocks, facilitating a smoother transition between highlighted areas and adjacent regions and reducing the influence of uneven illumination in recovering details. [ABSTRACT FROM AUTHOR]